seq: add an arithmetic ramp generator, and correct rep's length_out docs

[bobjansen]

rep tiles a pattern but never advances, so no form of it produces 0,1,2,...,N-1. Until now a row index cost two statements and a junk column: transforms require a materialized column name, so cumsum(rep(1)) inline fails with "cumsum: argument must be a column name" and you had to bind one = rep(1) first.

Table(1000)[update { i = seq(), t = seq(0.0, 0.5) }]

seq(from, by) fills that gap as a Generator, with no parser work: a FnKind::Generator row, a GeneratorExec entry, REPL completion and help. Length is always the frame's row count. Row i is computed as from + by*i rather than accumulated, so the float form does not drift.

Two things are deliberate and easy to get wrong later:

from and by are positional. infer receives only positional argument types, so a named from=0.0 would type the column Int64 and then fill it with Doubles; the kernel refuses named arguments rather than mistype. The names are also reserved keywords (KeywordFrom/KeywordBy), so seq(from=0) is a parse error and the named form cannot be written at all — the tests assert that through parser::parse, since require_ir REQUIREs a lower.

seq gets no effects.cpp entry despite sitting beside rand_* in every other registry: it is deterministic, and kEffNondet would wrongly block folding and CSE. rep is absent there for the same reason.

Also corrects SPEC 12.8: length_out was documented as if it set the output length, but N is the only value it ever accepts — anything else errors with "generates X rows but the frame has Y", in select and update alike. It is retained only for R fidelity. The docs no longer advertise the -1 sentinel or imply the length is overridable.

repl.cpp's completion array is reflowed by clang-format, since inserting one element re-pairs every subsequent line.